Biophysical Processor

The proposed algorithm is based on methods that have already been proven to be efficient.

They have been implemented to generate biophysical products from VEGETATION, MERIS, SPOT, and LANDSAT sensors. It mainly consists in generating a comprehensive database of vegetation characteristics and the associated SENTINEL2 or LANDSAT8 top of canopy (TOC) reflectances. Neural networks are then trained to estimate the canopy characteristics from the TOC reflectances along with set corresponding angles defining the observational configuration.

For the Sentinel 2 Products 2 different neural network architecture have been implemeneted, the NNET 20m and the NNET 10m. The two use a different combination of input bands as shown in the table below as input for the neural network in addtition to the auxiliary bands cos(viewing_zenith), cos(sun_zenith), cos(relative_azimuth_angle).

Band Central Wavelength (nm) Resolution (m) NNET 20m NNET 10m

B1 443 60

B2 490 10

B3 560 10 x x

B4 665 10 x x

B5 665 10 x

B6 740 20 x

B7 783 20 x

B8 842 10 x

B8a 865 20 x

B9 945 60

B10 1375 60

B11 1610 20 x

B12 2190 20 x

Band	Central Wavelength (nm)	Resolution (m)	NNET 20m	NNET 10m
B1	443	60
B2	490	10
B3	560	10	x	x
B4	665	10	x	x
B5	665	10	x
B6	740	20	x
B7	783	20	x
B8	842	10		x
B8a	865	20	x
B9	945	60
B10	1375	60
B11	1610	20	x
B12	2190	20	x

The NNET 10m uses only 10m bands and so it produce a 10m resolution output product, however it is capable of computing only the LAI, FAPAR and FVC indexes.

For the LANDSAT8 the same approach is used, but in this case the bands used as inputs are the green, red, near_infrared, swir_1 and swir_2 in addition to the auxiliary bands (zenith and azimuth both for sun and view).
In this case, as for the NNET 10m the algorithm is capable of computing only the LAI, FAPAR and FVC indexes.

Both NNET 10m, NNET 20m and LANDSAT8 neural networks are composed by three layers:

one input layer: made of 11 normalized input data for the NNET 20m, 6 for the NNET 10m and 7 for the LANDSAT8 one.
one hidden layer with 5 neurons with tangent sigmoid transfer functions
one output layer with a linear transfer function

In the final implementation there are two NNET 20m, one trained with S2A data and one with S2B data, and two NNET 10m, again trained for S2A and S2B respectively. This results in output more relaible and with better accuracy and take in accound the small differences in the reflectance response of the different sensors.

The actual algorithm running in SNAP runs the prediction step of the neural network, from the set of precomputed coefficients computed during the training phase.

For more details, please refer to the algorithm theoretical based document: